Q: how to calculate features on rolling windows but with fixed start point for each window?

[saheel1115]

Hi folks,

Firstly, grateful to you for creating this library. It has crucial features missing in other libraries and I love it so far.

I seek your help in figuring out:

• how to calculate features on entire time series?
• how to move the end point of window forward by a fixed amount (say, 1 week) but not move the start of the window? Basically, my data is weekly and I am solving a forecasting problem. I want to calculate time series features for each training sample without data leakage. Thus, for each week, I want to calculate features for the time series till that week; then move on to next week and calculate features for time series till that week; and so on.

Data:

Index	Sales
2023-01-02	4
2023-01-09	3
2023-01-16	7
2023-01-23	5

Output should be something like this:

Features considering only 1st week of data:

Index	Sales	Length	MeanSales
2023-01-02	4	1	4

Features considering first 2 weeks of data:

Index	Sales	Length	MeanSales
2023-01-09	3	2	3.5

Features considering first 3 weeks of data:

Index	Sales	Length	MeanSales
2023-01-16	7	3	4.67

Features considering first 4 weeks of data:

Index	Sales	Length	MeanSales
2023-01-23	5	4	4.75

Thanks in advance,
Saheel.

[saheel1115]

Update:

I am currently using this solution to achieve what I want:

• use a really large window size such that it covers the entire dataset
• use the index of DF as segment_end_idxs

fc = FeatureCollection(
    MultipleFeatureDescriptors(
          functions=[np.max, np.mean],
          series_names=["Sales"],
          windows="700days",
          strides="7days",
    )
)

fc.calculate(
    data=[df],
    include_final_window=True,
    segment_end_idxs=df.index[1:]
)

Is there a better solution? 😃